Semantic Error Detection Using ANN Feature Combinations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting semantic errors in text, such as semantic distance, confusing sets, and machine learning approaches, fail to achieve a desired degree of accuracy, particularly with semantic errors that involve words with similar grammatical properties used in the wrong context.
Innovation Solution
A system using an Artificial Neural Network (ANN) that generates features from statistical properties and co-occurrence information of words, combining N-gram features, word-word co-occurrence, and sentence-word co-occurrence matrices through matrix factorization, to detect semantic validity by learning mappings from feature sets to labels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional spelling checkers are used, then non-word spelling errors can be detected easily, but real-word semantic errors cannot be detected
Solution Approach 1:
The patent combines multiple detection methods (semantic distance method, confusing sets method, and machine learning classifier) into a unified system. The semantic distance method computes meaning degrees between words, the confusing sets method targets commonly mistaken word groups, and the machine learning classifier integrates these approaches with trained models to detect both non-word and real-word semantic errors comprehensively
2Measurement precision
If semantic distance method is used, then contextual meaning can be analyzed, but word sense disambiguation issues arise
Solution Approach 1:
The patent introduces an intermediary mechanism - the confusing sets method - that groups words commonly mistaken for one another. This intermediary approach simplifies the disambiguation process by pre-organizing problematic word pairs and triples, allowing the system to handle word sense disambiguation more efficiently while maintaining contextual analysis capabilities
3Measurement precision
If confusing sets method is used, then commonly mistaken words can be detected, but the approach fails when set size increases
Solution Approach 1:
The patent transitions from a two-dimensional confusing sets approach to a three-dimensional solution by integrating machine learning classifiers trained on large corpora. This adds a new dimension of statistical learning that handles large and dynamic word sets efficiently, overcoming the scalability limitations of traditional confusing sets methods while maintaining detection accuracy
4Measurement precision
If machine learning classifiers are used, then semantic errors can be detected using trained models, but training data requirements increase
Solution Approach 1:
The patent creates a universal machine learning classifier system that can detect multiple types of semantic errors (non-word errors, real-word errors, contextual errors) using a single trained model. The classifier is designed to handle diverse error patterns through unified feature extraction and classification mechanisms, reducing the need for separate training data for each error type while maintaining comprehensive detection capabilities
Data Source
AI summary
The invention provides a method and system for automatically detecting semantic errors in a text. In order to detect the semantic errors in the text, the method generates a first set of features and a second set of features corresponding to each word in the text. A first set of features corresponds to statistical features of a word and a second set of features for a word corresponds to the features generated for a plurality of words surrounding the word in the text using matrix factorization techniques. The method, then, combines a first set of features and a second set of features to generate a third set of features corresponding to each word in the text. Thereafter, the method utilizes an Artificial Neural Network (ANN) model to detect semantic validity of each word based on a corresponding third set of features.